Work log of a 2011 Mercedes-Benz G55 AMG (W463, M113K 5.4 L supercharged V8, 722.6 five-speed) and comma.ai's openpilot driving models. Since December 2025 both CAN buses have been decoded, comma's driving model put on the truck's own cameras, the transmission moved to open-source firmware and the engine ECU patched to take a torque request.
The truckselect a number
Perception and modelsControl and powertrainElectrical and CANaddedfactoryCAN-CCAN-B
What is on itall parts
| # | Part | Hardware | Software, rate | Since |
|---|---|---|---|---|
| 1 | Cameras added | ELP AR0234 on an 8 mm lens; Arducam B0576 IMX662 | MJPEG, 20 fps recorded; one case on the sun-visor mount | 10-22 |
| 2 | Tablet added | Galaxy Tab Active5 Pro | dashboard v139, Kotlin and Compose; voice assistant | 11-05 |
| 3 | Gateway, Raspberry Pi 5 added | Raspberry Pi 5 on NVMe | Rust gateway; 62 signals decoded live; every frame on both buses | 10-26 |
| 4 | Jetson Orin NX 16 GB added | Jetson Orin NX 16 GB, 2 TB NVMe | N16, TensorRT fp16, 32.5 ms a step; YOLO26s at 10 Hz; SegFormer-B0 at 5 Hz | 10-26 |
| 5 | Compute battery added | Eco-Worthy 314 Ah LiFePO4, 4,019 Wh; Victron Orion-Tr 12/12-30A | battery and charger read over Bluetooth every 10 s | 10-26 |
| 6 | Engine ECU, Bosch ME 2.8.1 factory | Bosch ME 2.8.1, Infineon C167, 512 KB flash | stock code and calibration; torque request at 50 Hz, 100 ms watchdog; 10 Hz telemetry | 11-03 |
| 7 | Transmission, 722.6 five-speed factory | 722.6 with the Ultimate-NAG52 controller | firmware V15: upstream dev with 5 G55 patches; live data at 10 Hz | 10-18 |
| 8 | Cruise control stalk factory | factory stalk on the steering-column module | the gateway's cruise controller; SET+ accepted 7 of 7 | 10-29 |
| 9 | Steering factory | factory hydraulic steering box, no motor | nothing steers yet; simulation only | factory |
| 10 | CAN buses factory | CAN-C 500 kbit/s, CAN-B 83.3 kbit/s | 131 signals in 47 messages in the DBC files; 62 decoded live | factory |
From the as-built record. Dates are 2026.
comma's releases on the truck's drivesnewest first; nothing ranked; newest f22e06e5
Against the release before it, paired on the same windows, interval clear of zero: ▼ lower▲ higherno separable changefirst on these windows. Every column is an error or a distance: lower is closer to what the truck did.
| Release | First public | Family | Params, M | Where | OL lateral error m/s², mean [95%] | CL path at 2 s cm, mean [95%] | CL path at 7.2 s m, mean [95%] | 2° heading, 2 s cm, mean [95%] | 0.5 m offset, 2 s cm, mean [95%] |
|---|---|---|---|---|---|---|---|---|---|
| f22e06e5 | 2026-11-05 | residual | 382.9 | Hugging Face | 0.18 [0.15, 0.19] | 14.1 [10.9, 17.1] | 2.46▲ [1.66, 3.46] | 48.1 [39.1, 59.6] | 25.4 [22.8, 28.1] |
| f22b8b34 | 2026-10-29 | residual | 382.9 | Hugging Face | 0.18▲ [0.15, 0.19] | 14.9 [11.4, 17.8] | 2.03▼ [1.36, 2.94] | 48.3▼ [39.2, 59.5] | 26.3 [23.8, 28.6] |
| 6b5b705f | 2026-10-17 | residual | 382.9 | Hugging Face | 0.16▼ [0.14, 0.17] | 13.9 [10.8, 17.2] | 3.01▲ [2.03, 4.08] | 50.0 [40.7, 61.5] | 26.7 [23.9, 29.3] |
| N16 nav model 96244b1f | 2026-10-15 | residual | 393.67 | Hugging Face | 0.17▲ [0.14, 0.18] | 13.6 [10.0, 17.9] | 2.40 [1.50, 3.42] | 48.4▼ [38.5, 61.4] | 25.8▼ [22.8, 29.0] |
| MD v3 / Flip Flop 1365126a | 2026-10-14 | residual | 382.9 | master | 0.16▼ [0.13, 0.17] | 13.7 [10.9, 16.6] | 2.53▲ [1.66, 3.59] | 51.7▲ [42.5, 62.9] | 28.5▲ [25.9, 31.0] |
| Mountain Dew 8d736683 | 2026-10-10 | residual | 382.9 | master | 0.16▲ [0.13, 0.18] | 13.2 [10.2, 15.8] | 2.05 [1.38, 3.04] | 49.2▼ [39.9, 60.5] | 25.8▼ [23.4, 28.2] |
| 870a4823 | 2026-10-08 | residual | 382.9 | Hugging Face | 0.16▼ [0.13, 0.17] | 13.3▼ [10.7, 15.4] | 2.04 [1.38, 3.01] | 51.8 [42.3, 63.4] | 28.7 [26.3, 31.1] |
| Cinque v3 f78ed37d | 2026-09-18 | big-rl | 381.59 | master | 0.17▲ [0.14, 0.18] | 15.0▲ [12.4, 17.2] | 1.87 [1.44, 2.35] | 51.0▲ [41.4, 62.7] | 28.2▲ [25.2, 31.0] |
| Cinque v2 1a421175 | 2026-09-11 | big-rl | 381.59 | master | 0.16 [0.14, 0.17] | 13.3▼ [10.7, 15.6] | 1.94 [1.45, 2.48] | 50.1▲ [40.4, 62.2] | 26.9 [24.3, 29.4] |
| lebowski | 2026-07-01 | lebowski | 876.48 | master | 0.17▼ [0.14, 0.18] | 17.0 [13.7, 20.3] | 1.98 [1.44, 2.54] | 45.3 [38.8, 53.7] | 26.7 [22.5, 30.2] |
| WMI | 2026-01-19 | vision-policy | master | 0.23▲ [0.20, 0.26] | not run | not run | not run | not run | |
| Dark Souls 2 | 2025-12-12 | vision-policy | master | 0.17 [0.14, 0.19] | not run | not run | not run | not run | |
| The Cool People's | 2025-10-20 | vision-policy | master | 0.17 [0.14, 0.18] | not run | not run | not run | not run | |
| North Nevada | 2025-10-09 | vision-policy | master | 0.17 [0.14, 0.17] | not run | not run | not run | not run | |
| Steam Powered | 2025-08-17 | vision-policy | master | 0.16▼ [0.14, 0.17] | not run | not run | not run | not run | |
| Filet o Fish | 2025-03-07 | vision-policy | master | 0.20▼ [0.16, 0.21] | not run | not run | not run | not run | |
| North Dakota | 2024-04-29 | supercombo | master | 0.21 [0.17, 0.23] | not run | not run | not run | not run | |
| Certified Herbalist | 2024-02-19 | supercombo | master | 0.22 [0.17, 0.23] | not run | not run | not run | not run | |
| Farmville | 2023-11-07 | supercombo | master | 0.21 [0.18, 0.22] | not run | not run | not run | not run |
open loop: 119 windows from 12 drives; closed loop: 98 windows from 13 drives; recovery: 100 windows from 14 drives. Intervals resample whole drives; the comparison is paired on the same windows. Hover a cell for the difference. results/<id>.json for every number.
FindingsPerception and modelsControl and powertrainElectrical and CAN 8 notes
| Date | Finding | Result | n | Grade |
|---|---|---|---|---|
| 11-04 | N16's frozen vision tokens do not replace a box detector | 2.7 % of vehicles, against 55.5 % for YOLO26s | recorded drives, offline | measured |
| 11-04 | Camera position moves where N16 holds the truck in its lane | -0.97 m per m of camera shift | 135 windows | simulation |
| 11-04 | A trained steering correction on N16 recovers more often in simulation | +0.18 [+0.13, +0.33] recovery at 1.5 m | 44 windows, 3 drives | simulation |
| 11-03 | The camera loop ran below 20 Hz because OpenBLAS threads spun | 17.7-18.8 Hz, then 19.91-19.92 Hz | 3 drives after the fix | measured |
| 10-31 | comma's models turn the wrong way first because of the base policy | 258 of 258 | pushes | simulation |
| 10-30 | After a push, N16 still encodes the truck's offset and heading | R² 0.84-0.93 from the tokens | held-out sessions | simulation |
| 10-29 | Teaching comma's world model to keep a sideways offset | road gain 1.05 at 3 s; fine-tuned 1.24 | small test sets | simulation |
| 10-29 | A truck model learned from the G55's CAN predicts motion better | 0.29 m after 4 s, against 7.34 m | 12 held-out drives | measured |
Open questionswhat decides the next step
| Question | Blocks | Settled by |
|---|---|---|
| Does the ESP take SET−, RESUME and CANCEL from the gateway? | full cruise control | a parked test on the truck |
| Does the ESP relay a braking request from the gateway? | braking with no new hardware | a parked test, then a road test |
| How hard does factory cruise brake, and does it fade on a descent? | longitudinal safety | measured descents |
| Does a donor electric column fit between the steering column and the centre joint? | the steering actuator | one measurement on the truck |
| What torque does the steering take at the rim, stationary and rolling? | the steering torque budget | a measurement on the truck |
| Is the closed-loop renderer valid at night? | every night result | the world model's geometry check, day and night apart |
| Does a reward-trained N16 recover from pushes more often on held-out drives? | training inside comma's model | the locked test and a fresh evening |
Logall
| Date | Kind | What |
|---|---|---|
| 11-06 | test change | The site is renamed My Self-Driving G-Wagon and moves to https://my-self-driving-g-wagon.drose.io; openpilot-gwagon-evals.drose.io redirects there. |
| 11-06 | test change | The Signals page lists every message in the DBC files, the steering-column message the gateway sends for cruise control included; the build no longer refuses command frame IDs. |
| 11-06 | test change | The site is dense: one front page with the truck's parts and what is on each, every release's G55 results with intervals in one table, the findings, the open questions, the log, the signals and a glossary. |
| 11-06 | test change | The site is a log of the G55 work. |
| 11-05 | release | f22e06e5 uploaded to Hugging Face |
| 11-05 | tested | 1 model scored, suite geval-1 |
| 11-05 | test change | The site follows the releases over time instead of reporting one snapshot: a timeline of every release by model family, a feed of releases, test runs and test changes (also as RSS), one page per test with every release's 95% interval, newest first, and one page per release. |
| 11-04 | note | N16's frozen vision tokens do not replace a box detector |
| 11-04 | note | Camera position moves where N16 holds the truck in its lane |
| 11-04 | note | A trained steering correction on N16 recovers more often in simulation |
| 11-03 | note | The camera loop ran below 20 Hz because OpenBLAS threads spun |
| 10-31 | note | comma's models turn the wrong way first because of the base policy |
| 10-31 | tested | 9 models scored, suite geval-1 |
| 10-31 | test change | The site is a plain results report: one table per test, each model's mean with its 95% interval and its median, newest release first. |
Signalsall 131
| Sender | Bus | Signals |
|---|---|---|
| ESP | CAN-C | 18 |
| SAM_V | CAN-B | 16 |
| SAM_H | CAN-B | 15 |
| EZS | CAN-B | 13 |
| TCU | CAN-C | 11 |
| ECU | CAN-C | 9 |
| MRM | CAN-C | 9 |
| UNKNOWN | CAN-C | 8 |
| DOOR_FL | CAN-B | 7 |
| ZGW | CAN-B | 5 |
82 measured; the rest not yet labelled in the DBC. One truck, not a Mercedes specification.
Since December 20253,154 commits
- The engine ECU runs patched firmware that accepts a torque request on CAN.
- The gateway drives the factory cruise control over CAN.
- Every new comma driving model is tested on the truck's drives as it is released.
- The transmission runs a custom Ultimate-NAG52 firmware build; the object and free-ground models run beside the driving model.
- A closed-loop simulator built from the truck's recorded drives.
- comma's driving model runs on the Jetson at 20 Hz on the truck's camera.
- The Jetson's first live camera detections, YOLO26 on TensorRT.
- The in-car gateway designed: a Raspberry Pi on both CAN buses and the engine ECU's K-line.
- CAN-B identified as fault-tolerant CAN and the bus adapter chosen; the Ultimate-NAG52 transmission controller ordered.
- The project's repository starts.
Release timelineevery comma release by model family
shipped in an openpilot releaseon openpilot masteron Hugging Face onlynewestnot scored here
Swipe the timeline sideways for earlier releases.
Glossary
N16 comma's driving model 96244b1f_12800 (N16 nav), 406M parameters, with navigation inputs. The model on the truck.
push the simulated truck moved 1.0 or 1.5 m sideways off its recorded path; the model must bring it back.
recovery the share of pushes the model returns from: back within 0.3 m for 1 s, inside a safe corridor, still moving.
bench8 the closed-loop simulator built here: the model steers a simulated G55 while the camera view is redrawn from recorded frames.
open loop (OL) the model watches a recorded drive; its requests are compared with what the truck did next.
closed loop (CL) the model drives a simulated G55 in comma's world model; scored by distance from the driver's recorded path.
worldmodel-4B comma's video model that draws the next camera frames for a given motion.
residual a small learned correction added to a frozen model's steering output.
window a stretch of a recorded drive used as one test case. Intervals resample whole drives.
KOEO key on, engine off: ignition in position 2.
CAN-C, CAN-B the W463's powertrain and chassis bus (500 kbit/s) and its fault-tolerant body bus (83.3 kbit/s).
Ultimate-NAG52 an open-source replacement for the 722.6 transmission's EGS52 controller.